Research And Application Of Circular Convolution Parallel Extreme Learning Machine

In the process of solving practical problems,the nonlinear complex system with large delay and strong coupling is often encountered,making a problem that it is difficult for the traditional mechanism modeling to build an accuracy mathematical model.But the emergence of neural networks provides an effective method to solve the above problem.However,in the application process,classical neural networks decide its weight value and bias value though the random gradient descent method,causing the shortcomings of too long training time and easing to fall into the local optimal.In response to above problems,extreme learning machine(ELM)obtains its unique learning mechanism and possesses very good generalization ability,overcoming the weakness of the traditional network.Therefore,a lot of work on ELM and its improvements have been done in this research,and then its improved algorithms are applied to the UCI datasets and practical engineering modeling.The main content of this paper is described as follows:(1)Based on ELM,a new neural network algorithm called Circular Convolution Parallel Extreme Learning Machine(CCPELM)is proposed.And then nine classic UCI Benchmark regression problems have been applied to verify CCPELM's validity.Then,in order to further improve the performance of CCPELM,CCPELM is improved and then we call it Upgrades Fast Learning Network(U-FLN).Through the above problems,U-FLN shows better generalization ability than other excellent algorithms.In addition,in order to verify the effectiveness of U-FLN in practical application,CCPELM and U-FLN are applied to build the thermal efficiency model of a 300 MW pulverized coal fired boiler.The simulation results show that U-FLN model can better predict the thermal efficiency of boiler than other models.(2)Based on the above off-line algorithm,an online learning algorithm called Online Sequential Circular Convolution Parallel Extreme Learning Machine(OCCPELM),is proposed.OCCPELM is applied to the modeling of six UCI datasets,and the simulation results show that OCCPELM with "sigmoid" activation function can obtain smaller prediction error than OSELM,and it obtains good stability in a big range number ofhidden layer node.Then,the online algorithm is still applied to the on-line simulation of the thermal efficiency of a 300 MW pulverized coal fired boiler.The experiment results show that OCCPELM model has better generalization ability than other excellent models and has good stability and quick learning ability.More importantly,compared with other models,the OCCPELM model is more sightly affected by the number of initial training samples,and it could be considered as a candidate to accomplish the online modeling and the prediction of thermal efficiency.